Urdu-Llama-3.2-3B-Instruct-v1 Overview

Developed by Khurram Pervez (Khurramcoder), this model is a specialized fine-tuned version of Meta's Llama-3.2-3B-Instruct, designed for advanced Urdu language processing. With 3.2 billion parameters and a 32768 token context length, it focuses on delivering high-quality instruction following and text generation in Urdu.

Key Capabilities

• Native Urdu Reasoning: The model was trained on the large-traversaal/urdu-instruct dataset (51.7k rows), enabling it to perform complex tasks such as translation, creative writing, and question answering with a deep understanding of Urdu cultural nuances.
• Efficient Architecture: Fine-tuned using Unsloth and QLoRA, this model achieves powerful performance while maintaining a lightweight footprint, making it efficient for deployment.
• Optimized for Urdu Script: It incorporates the latest Llama 3.2 multilingual tokenizer, which significantly improves its handling and generation of Urdu script.

Good For

• Urdu Instruction Following: Excels at understanding and executing instructions provided in Urdu.
• Urdu Text Generation: Capable of generating high-quality, culturally relevant Urdu text for various applications.
• Translation and QA: Particularly strong in translation tasks involving Urdu and answering questions in Urdu, leveraging its specialized training data.